[squid-users] Squid performance not able to drive a 1Gbps internet link

Amos Jeffries squid3 at treenet.co.nz
Wed Aug 3 03:30:37 UTC 2016


On 3/08/2016 2:42 p.m., Heiler Bemerguy wrote:
> 
> in /etc/sysctl.conf, add:
> 
> net.core.rmem_max = 8388608
> net.core.wmem_max = 8388608
> net.core.wmem_default = 32768
> net.core.rmem_default = 32768
> net.ipv4.tcp_wmem = 1024 32768 8388608
> net.ipv4.tcp_rmem = 1024 32768 8388608
> 


PLease aso bump up your version to 3.5.20, there have been more than a
few performance bug fixes since 3.5.1.

For near-Gbps speeds you will need one of the 'Extreme CARP'
configurations for spreading the workload between Squid processes. That
has achieved 800-900 Mbps a while back.
<http://wiki.squid-cache.org/ConfigExamples/ExtremeCarpFrontend>

If thats not fast enough, you may also wish to patch in a larger value
for HTTP_REQBUF_SZ in src/defines.h to 64KB with a matching incease to
read_ahead_gap in squid.conf. That has had some mixed results though,
faster traffic, but also some assertions being hit.

You may find that memory becomes your bottleneck at higher speeds.
8-16GB sounds like a lot for most uses, but when you have enough
connections active to drive Gbps (with 4-6x 64KB I/O buffers) there are
is lot of parallel pressures on the RAM.

Amos



More information about the squid-users mailing list